Created by: NHZlX

We set the same dims of Multihead matmul op 's input to output original.

eg:

 auto dim_input = context->GetInputDim("Input");
 context->SetOutputDim("Out", dim_input);
 context->ShareLoD("Input", /*->*/ "Out");

https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/operators/fused/multihead_matmul_op.cc#L80

The setting above is wrong.

Let's first introduce some noun variables in Multihead matmul op:

1) Batch : the batch size.
2) SeqLen: the sequence length.
3) HiddenSize: the second dims size of the Embedding table.
4) HeadNumber: the head number of the encoder.
5) HeadSize: the size of each Head.

The Multihead matmul op's Input Dims is [Batch * SeqLen * HiddenSize], the Output Dim should be [Batch * SeqLen * (HeadNumber * HeadSize)].